Triple points: From non-Brownian filtrations to harmonic measures (Q1380469)

Two seemingly unrelated problems are solved in this lengthy paper full of wonderful ideas: In 1979 \textit{M. Yor} asked the following question: (1) If \(B\) is a \(({\mathcal F}_t)\) Brownian motion having the predictable representation property with respect to \(({\mathcal F}_t)\), is \(({\mathcal F}_t)\) the natural filtration of a Brownian motion? (2) Let \(U\) be a bounded domain in \({\mathbb{R}}^d\). The harmonic boundary of \(U\) is defined as a measure type on \(\overline{U}\setminus U\). For each \(d\geq 1\) there is a finite \(N_d\) such that if \(U_1, \dots, U_{N_d+1}\) are pairwise disjoint bounded domains, then the intersection of their harmonic boundaries is empty. It was conjectured in \textit{C. Bishop} [Ark. Mat. 29, No. 1, 1-24 (1991; Zbl 0733.31005)], that \(N_d=2\) for all \(d\) (same as in the case of smooth Euclidean boundaries). The conjecture is proved in this paper. Problem (1) was solved negatively in \textit{L. Dubins, J. Feldman, M. Smorodinsky} and \textit{B. Tsirelson} [Ann. Probab. 24, 882-904 (1996; Zbl 0870.60078)], by constructing a rather complicated counterexample. It has been conjectured for some time that if \(Z\) is a Walsh's Brownian motion, then its natural filtration \(({\mathcal F}_t^Z)\) is not generated by a Brownian motion. Since it is known that \(({\mathcal F}_t^Z)\) has a predictable representation property, proving this conjecture would give a simple counterexample to problem (1). A proof of this conjecture is the first main result of the paper under review. In order to prove the result, Tsirelson introduced a new invariant of filtrations taking two values - ``cozy'' and ``non-cozy'', and showed that a filtration generated by a finite or infinite collection of independent Brownian motions is cozy. The next step was to show that the filtration \(({\mathcal F}_t^Z)\) of Walsh's Brownian motion \(Z\) is non-cozy. The proof relies on the idea that if \(({\mathcal F}_t^Z)\) were cozy, then \(Z\) and its \(\rho\)-correlated copy \(Z^{\rho}\) would have many common zeros. On the other hand, if \(R_1\) and \(R_2\) are \(\rho\)-correlated copies of a reflecting Brownian motion, then \(\int_0^t 1_{(R_1(s)=0)}dL_s(R_2)=0\) (where \(L(R_2)\) is the local time of \(R_2\)), showing that common zeroes of \(R_1\) and \(R_2\) (and therefore of \(Z\) and \(Z^{\rho}\)) are rare. In the sequel of the paper the author proves a version of the above result robust under change of measure. More precisely, if \((\Omega, {\mathcal F}^Z, P)\) is a probability space generated by Walsh's Brownian motion \(Z\), and \(Q\) is a probability measure equivalent to \(P\), then \((\Omega, {\mathcal F}^Z, Q)\) is non-cozy. Results similar to the described ones were also proved in \textit{M. T. Barlow, M. Émery, F. B. Knight, S. Song} and \textit{M. Yor} [Autour d'un théorème de Tsirelson sur des filtrations browniennes et non browniennes, in Séminaire de Probabilités XXXII, Lect. Notes Math. 1686, 264-305 (1998)]. In order to deal with problem (2), another generalization of the above result was needed. Let \({\Sigma}_+\) denote the set of nonnegative continuous semimartingales on \((\Omega, {\mathcal F}, P)\) of the form \(X=M+V\) with \(dV\) carried by \(\{t: X_t=0\}\). If \(X^{(1)}, X^{(2)}, X^{(3)} \in {\Sigma}_+\) are nonoverlapping (i.e., at most one is non zero), and if the filtration is cozy, then \(dV^{(1)} \wedge dV^{(2)} \wedge dV^{(3)} = 0\). This result leads to the following theorem which establishes the conjecture in (2): Let \(U_1, U_2, U_3\) be pairwise disjoint bounded domains in \({\mathbb R}^d\), \(d\geq 1\), \(x_k\in U_k\), \(k=1,2,3\), and let \({\mu}_k={\mu}_{U_k,x_k}\) be the harmonic measure on \(\partial U_k\), \(k=1,2,3\). Then \({\mu}_1 \wedge {\mu}_2 \wedge {\mu}_3 =0\).